The combination of forecasts resulting from an ensemble of neural networks has been shown to outperform the use of a single "best" network model. This is supported by an extensive body of literature, which shows that combining generally leads to improvements in forecasting accuracy and robustness, and that using the mean operator often outperforms more complex methods of combining forecasts. This paper proposes a mode ensemble operator based on kernel density estimation, which unlike the mean operator is insensitive to outliers and deviations from normality, and unlike the median operator does not require symmetric distributions. The three operators are compared empirically and the proposed mode ensemble operator is found to produce the most accurate forecasts, followed by the median, while the mean has relatively poor performance. The findings suggest that the mode operator should be considered as an alternative to the mean and median operators in forecasting applications. Experiments indicate that mode ensembles are useful in automating neural network models across a large number of time series, overcoming issues of uncertainty associated with data sampling, the stochasticity of neural network training and the distribution of the forecasts.

No abstract

Inventory control systems rely on accurate and robust forecasts of future demand to support decisions such as setting of safety stocks. Combining forecasts is shown to be effective not only in reducing forecast errors, but also in being less sensitive to limitations of a single model. Research on forecast combination has primarily focused on improving accuracy, largely ignoring the overall shape and distribution of forecast errors. Nonetheless, these are essential for managing the level of aversion to risk and uncertainty for companies. This study examines the forecast error distributions of base and combination forecasts and their implications for inventory performance.

Forecast selection and combination are regarded as two competing alternatives. In the literature there is substantial evidence that forecast combination is beneficial, in terms of reducing the forecast errors, as well as mitigating modelling uncertainty as we are not forced to choose a single model. However, whether all forecasts to be combined are appropriate, or not, is typically overlooked and various weighting schemes have been proposed to lessen the impact of inappropriate forecasts. We argue that selecting a reasonable pool of forecasts is fundamental in the modelling process and in this context both forecast selection and combination can be seen as two extreme pools of forecasts. We evaluate forecast pooling approaches and find them beneficial in terms of forecast accuracy. We propose a heuristic to automatically identify forecast pools, irrespective of their source or the performance criteria, and

Recently, combination algorithms from machine learning classification have been extended to time series regression, most notably seven variants of the popular AdaBoost algorithm. Despite their theoretical promise their empirical accuracy in forecasting has not yet been assessed, either against each other or against any established approaches of forecast combination, model selection, or statistical benchmark algorithms. Also, none of the algorithms have been assessed on a representative set of empirical data, using only few synthetic time series. We remedy this omission by conducting a rigorous empirical evaluation using a representative set of 111 industry time series and a valid and reliable experimental design. We develop a full-factorial design over derived Boosting meta-parameters, creating 42 novel Boosting variants, and create a further 47 novel Boosting variants using research insights from forecast combination. Experiments show that only few Boosting meta-parameters increase accuracy, while metaparameters derived from forecast combination research outperform others.